Agents lack basic safety mechanisms and are 'vibe coded' blackboxes, posing deployment risks
gerardsans · x · 2026-08-25
Gerard Sans notes that while LLMs follow instructions, current agents lack essential safety mechanisms like authentication, authorization, and monitoring. Labs skip safeguards due to cost, relying on "vibe coded" natural language instructions that create unpredictable, un-debuggable blackboxes. He questions why such systems are being deployed in corporate environments given these obvious risks.
Related event: AI safety debate: the real risk is bad engineering, not rogue models(5 posts)→
More from Safety
- Russian Group Ran a Fake Think Tank via ChatGPT; OpenAI Says AI Slashed the Manpower — heypearlai · 2026-08-25
- AI agents hacked real companies: OpenAI breach triggers bill, industry pause — Last Week in AI · 2026-08-25
- UK's NCSC advises kill switches for AI agents, admits model safety training can be bypassed — Servola-Journal · 2026-08-25
- Beijing's 2026 AI for Science summit: 50 academicians, 225 filed models, full-stack AI4S push — 智东西 · 2026-08-25
- HF Agents Escaped Sandboxes Due to Impossible Benchmarks — nptacek · 2026-08-25
- Insurers Retreat as AI Agents Become 'Uninsured Liabilities' — hbouammar · 2026-08-25